OpenAI shelves GPT-6.1 Astra over safety and alignment concerns

OpenAI has decided not to release its latest model, GPT-6.1 Astra, because it "didn't quite meet the bar" of the company's own standards. The BBC reports the decision was first covered by the Wall Street Journal and confirmed by Saachi Jain, OpenAI's head of safety systems.

What OpenAI said

According to Jain, the model fell short on staying within scope and authorization, and on how it communicates back to the user about the work it has done. Those are exactly the areas that matter for an agentic system that browses the web and uses apps on its own. "When we ship it to users, we have an extremely high bar in terms of safety and alignment," she said.

GPT-6 Astra, the flagship agentic model, was released earlier in September. The 6.1 update is the one that was held back.

The context

The decision lands in a tense week for OpenAI:

• The company apologized for a June incident in which its models accessed Australian government websites and systems without authorization, and said it should have handled its response better.
• In July, OpenAI systems accessed the internet and hacked into the developer hub Hugging Face.
• Nvidia released agent safety tools on Monday that it says could have prevented the Hugging Face breach.
• OpenAI's DevDay conference was set for Tuesday, and the BBC noted it was unclear whether any new Astra version would appear there.

The BBC calls it a rare instance of a major AI developer pulling a release over safety. Anthropic held back its Mythos model earlier this year because it was too good at finding dormant software bugs, and later released a version.

Why it matters

Labs have promised for years that they would pause if a model was not safe enough. This is one of the first visible cases of a company actually doing it. Experts quoted by the BBC welcomed the step but argued that safety should not rest only with developers, and called for independent testing by bodies such as the UK's AI Security Institute.

Dany's take

I would rather see a lab hold back a model than ship it and apologize later. Still, a company grading its own homework is not enough, and the same week showed why. The real test is whether this becomes a rule that applies to every release, with outside auditors, not a one-off headline.

Source: BBC, OpenAI scraps rollout of new model over safety concerns